NVIDIA and Oracle expand Vera Rubin to multi-gigawatt OCI scale
NVIDIA and Oracle said they will expand Oracle Cloud Infrastructure Superclusters to multi-gigawatt scale with NVIDIA Vera Rubin NVL72 systems, while also making NVIDIA Vera Rubin NVL72 and NVIDIA Vera CPUs generally available on OCI. The companies said the added capacity is aimed at customer demand for AI training, inference and agentic applications, and at helping organizations scale CPU and GPU resources separately so more infrastructure can be kept in use.
Evidence and context
In the announcement dated Oct. 27, 2026, NVIDIA said the planned OCI expansion builds on existing NVIDIA-powered infrastructure at Oracle, including GB200 and GB300 NVL72 systems. Oracle said its AI infrastructure collaboration with NVIDIA now covers the GPU-heavy work of training and serving frontier models as well as CPU-intensive agent tasks between model calls, such as running tools, testing code and retrieving data.
The companies also said OCI plans to introduce NVIDIA HGX Rubin NVL8, an eight-GPU configuration that uses NVLink in an x86-based server design. According to the announcement, the planned system includes two 64-core Intel Xeon 6 CPUs, 4TB of DDR5 memory and 6.4 Tbps of RDMA networking. Oracle said this would provide a smaller deployment option for training, fine-tuning and inference while maintaining compatibility with x86-based applications.
Verified background material supplied with the release describes Vera Rubin NVL72 as a rack-scale system combining 72 Rubin GPUs and 36 Vera CPUs with ConnectX-9 SuperNICs and BlueField-4 DPUs. NVIDIA’s current specifications list 20.7 TB of aggregate HBM4 memory, 1,400 TB/s aggregate GPU-memory bandwidth and 216 TB/s of NVLink 6 rack-level bandwidth, or 3 TB/s per GPU. The same product table lists 3,600 PFLOPS of NVFP4 inference with sparsity, 2,520 PFLOPS of dense NVFP4 training, 1,260 PFLOPS of dense FP8/FP6 training and 288 PFLOPS of dense FP16/BF16. These are specification figures, not independently measured application throughput.
The release cites OpenAI as an example of the platform in use. It says OpenAI has trained and served GPT-6 Astra in Abilene on an OCI Supercluster with more than 100,000 NVIDIA Blackwell GPUs. OpenAI engineering leader Sreekanth Pothanis said NVIDIA Vera Rubin on OCI gives the company computing performance and scale to accelerate model training and serve more demanding AI workloads. OpenAI CTO of compute Uday Ruddarraju said Vera CPUs on OCI provide the loaded performance needed to execute code and run tools across growing agentic sandbox workloads.
The announcement does not provide a timetable for the multi-gigawatt Vera Rubin expansion or for the planned HGX Rubin NVL8 rollout. It also does not disclose pricing, customer counts beyond named examples, or financial terms tied to the expanded collaboration.